| According to GPC(global product classification)classification standard,there are thousands of commodity categories.Realizing fast and accurate automatic commodity classification can effectively reduce the burden of staff.Commodity classification can be classified by commodity pictures,commodity names and commodity description information.This paper classifies commodities by commodity names.Because commodity name belongs to text and is short in length,this paper focuses on improving the short text classification method and applying it to the task of commodity classification.This paper chooses BERT model to fine tune and applies BERT with other models to the task of commodity classification,and the main work is as follows:(1)Fine tune BERT and apply it to commodity classification task.This paper focuses on the structure and working principle of the BERT model.By fine tuning the BERT model for commodity classification task,and training the fine tuned BERT model on the self-made data set and the e-commerce data set obtained by the open research data platform of Peking University,it is concluded that the BERT model has excellent generalization ability in commodity classification task than other benchmark models.(2)This paper combines the BERT model with TextRCNN to solve the catastrophic forgetting problem of the fine tuned BERT model in the commodity classification task.TextCNN uses two-channel input.The first channel uses the word vector embedded by the word of BERT model as the input,and the word vector will not be updated with the training;The second channel uses the word vector obtained by Word2 Vec word embedding as the input,and the word vector as the parameter will be updated continuously with the training of the network,so as to make up for the information forgotten in the word embedding process of BERT model.Experiments show that the generalization ability of BERT-TextCNN is better than that of the finetuning BERT model.(3)This paper combines the BERT model with TextRCNN to solve the catastrophic forgetting problem of the fine tuned BERT model in the commodity classification task.TextRCNN adopts Bi LSTM.Because LSTM has long-term memory ability and can select the information of memory sequence in different states,it can solve the catastrophic forgetting problem of BERT model in commodity classification task.Experiments show that the generalization ability of BERT-TextRCNN model is better than Bert-TextCNN model.(4)Based on the Spring Boot framework,this paper introduces the BERTTextRCNN model into the actual automatic commodity classification task,designs and implements the automatic commodity batch classification system. |
